This AI cannot hold a conversation. Microsoft designed it that way
On 9 October Microsoft released Microsoft-Decision-1 in Foundry. You give it a question and a fixed set of answers. It does not write a paragraph back. It returns a probability for each answer. Ask "this agent wants to delete a folder, let it run?" and you get something like: allow 0.12, send to a human 0.88.
The goal is calibration: when it says 90%, it should be right about nine times out of ten.
Its jobs are the unglamorous ones: routing, classifying, prioritising, grading another AI's answer, and checking an agent's next action before it is allowed to run.
When I build agents, the part that takes the longest to tune is never how they talk. It is how they choose. Is this email worth waking me for? Can this step go ahead on its own? Those yes/no calls happen hundreds of times a day, and asking a large model every time is slow and expensive.
Microsoft's own numbers (no independent tests yet):
・Built on Alibaba's open-weight Qwen3.5-9B, post-trained by Microsoft
・About 35 times faster than GPT-6 Sol at median latency
・Xbox Research used it to sort more than 10,000 pieces of feedback: quality competitive with GPT-6 Sol, over 14 times faster, 200 times cheaper
・Input at $0.042 per million tokens, output free
The goal is calibration: when it says 90%, it should be right about nine times out of ten.
The part I would read first, though, is the catalogue page on how not to use it. It is not designed or evaluated to be the sole automated decision-maker in consequential decisions about people. Credit, employment, housing, insurance, education, healthcare and legal rights are each named. Thresholds, escalation and human oversight are the job of whoever builds it in.
One more line: it does not explain why it chose.
A model that gives no reasons, telling you in writing not to let it decide alone. I will take that.
If an AI could choose faster and cheaper than you but never told you why, what would you let it choose for you, and what would you never hand over? Tell me what you think in the comments 👇
Microsoft: https://commandline.microsoft.com/microsoft-decision-1-model-foundry/
More in AI news
All →
Two frontier models in two days. On 12 October Meta released Muse Wonder
From "there's an app for that" to "there's an app for me". That line is Google's own
Parental controls switched on. Does that mean you would hear about anything serious?
A country that teaches AI from kindergarten has just said the human part of school cannot…